US 20050251291 A1
A system of self-organizing mobile robotic agents (MRAs) in a multi-robotic system (MRS) is disclosed. MRAs cooperate, learn and interact with the environment. The system uses various AI technologies including genetic algorithms, genetic programming and evolving artificial neural networks to develop emergent dynamic behaviors. The collective behaviors of autonomous intelligent robotic agents are applied to numerous applications. The system uses hybrid control architectures. The system also develops dynamic coalitions of groups of autonomous MRAs for formation and reformation in order to perform complex task.
1. A multi-robotic system architecture having a plurality of system layers interconnected to one another, comprising:
a first layer including a hybrid control system for a plurality of mobile robotic agents (MRAs);
a second layer including a distributed mobile robotic system for the plurality of MRAs;
a third layer including a grid computing architecture in a plurality of dynamic clusters;
a fourth layer including a dynamic distributed object relational database management system;
a fifth layer including an omni-nodal evolutionary artificial neural network;
a sixth layer including a multi-agent system and a plurality of intelligent mobile software agents;
a seventh layer including a plurality of cellular automata simulations; and
an eighth layer including a plurality of functional applications.
The present application claims the benefit of priority under 35 U.S.C. §119 from U.S. Provisional Patent Application Ser. Nos. 60/404,945 and 60/404,946, filed on Aug. 21, 2002, the disclosures of which are hereby incorporated by reference in their entirety for all purposes.
There are several categories of prior art patents that apply to the present invention. These patents involve mainly mobile robots and groups of mobile robots.
Matsuda (robot system and control device), U.S. Pat. No. 5,825,981; Peless et al. (method for operating a robot), U.S. patent application publication number # 20010047231; and Nourbakhsh et al. (socially interactive autonomous robot), U.S. patent application publication number # 20020013641, mobile robots are used automatically, or with manual intervention to perform tasks such as multifunctional manufacturing, cleaning, mowing, snow blowing or interacting with humans. These pedestrian approaches to robotic control fit into the main paradigm of robotic applications.
Kawakami (mobile robot control system), U.S. Pat. No. 5,652,489; Asama et al., (mobile robot sensor system), U.S. Pat. No. 5,819,008; and Wallach et al. (autonomous multi-platform robot system), U.S. Pat. No. 6,374,155 involve multiple mobile robots. These patents involve using sensors for navigation and obstacle avoidance. In addition, one mobile robot can transmit information to another mobile robot for some effect. These inventions offer only rudimentary connections between robots and lack advanced system functions.
Most of the research history involving the technologies of the present system—including (1) intelligent agents and self-organizing systems, (2) AI and D-AI in coordinated systems, (3) negotiation and problem solving and (4) cooperating agents and aggregation—are represented in the academic literature, described below.
The development of complexity theory is fairly recent. Theorists from economics and biology advanced the view in the 1980s that systems are self-organizing and adaptive of their environments. In particular, biologists have studied ant and insect social organization and have observed the complex adaptive behaviors of these societies.
Researchers at the Sante Fe Institute (SFI) have developed complexity theory by looking at the fields of biology, economics, mathematics, epistemology and computer science. One of the aims of the SFI is to develop a complex self-organizing computer model representing artificial autonomous agents that emulate the biological functions of complex insect social behavior.
SFI theorists have developed the swarm intelligence model of artificial computer societies primarily for simulating economic systems. The swarm intelligence model, by emulating biological system operation, uses ideas of emergent behavior to describe the complex social interactions of relatively simple insects according to straightforward decentralized rules governing group activity.
The challenge for computer scientists lies in how to develop a system of self-organized autonomous robotic agents. The development of societies of behavior-based robotics that fuse elements of system control with elements of decentralized local control is one of the most difficult challenges in computer science and robotics. A key part of this problem lies in how to configure AI systems for problem solving in a MRS for collective behavior. In short, how can we design an intelligent MRS for optimal adaptation to dynamic environments?
The computer science fields of robotics and AI have evolved in the past decade in such a way that a convergence of technologies allows an explosion in research in collective robotics and in intelligent systems in order to achieve the goals of developing an intelligent MRS for group behavior. For example, rapid advances in computation resources, communications and networking allow the combination of integrated technologies necessary for a development of a sophisticated MRS. In addition, in the area of AI research, several trends have emerged, including GA, GP, A-NN and distributed AI, that allow computer systems to not only learn but achieve some degree of autonomy.
In the early 90s, Brooks developed a decentralized modular approach to robotics at MIT's Media Lab. Revolutionary at the time because it spurned conventional wisdom of highly computation-intensive deliberative robotic control approaches, his modular approach used less than three percent of traditional computer approaches. This leap in efficiency was achieved by separating the subsystems for automatic reactive control (he called it subsumption) rather than deliberative top-down robot system control. The mobility, navigation and pick-up functions of the robot could be separated for increased efficiency.
By exploiting this research stream, Arkin (1998) developed a behavior-based model of robotics. In this model, Arkin describes behavior-based robotic architectures as well as experiments in the field with sophisticated hybrid robotic architectures. An example of this hybrid approach is NASA's Atlantis system (1991) that synthesizes deliberative planning with group behavior. The aim of these models is to develop autonomous robots that are adaptive to their environment. The development of robotic teams with social behavior is one of the most difficult challenges, according to Arkin's pioneer study.
Bonabeau et al. (1999), an SFI fellow, develops a research stream that connects the study of ant and insect behavior in complex biological social systems with the development of complex artificial robotic societies. In their vision of swarm intelligence, they use key notions of system self-organization, reactive behavior and environmental adaptation to point to a model for artificial robotic systems that might emulate biological systems.
In 2001, Kennedy and Eberhart focused on the social and theoretical aspects of swarm intelligence. Their examination of group behavior develops a computer model of adaptive self-organized systems, similar to economic “particle” simulations by the SFI, by emulating the social behavior of biological systems. In order to develop an artificial swarm system, the authors look to complex pattern emergence, which has a lineage from Von Neumann to Burks to Wolfram. In this research stream, cellular automata are used to simulate a complex but stable self-organizing system. Though the authors refer to research experiments with robot societies, their focus remains on computer and theoretical models of complex social behavior involving autonomous entities.
Another important research stream involves the application of AI to networks. The emergence of the Internet has presented novel ways to conduct commerce automatically with autonomous software agents in a MAS. Originally developed by Smith, the contract-net protocol established an early model for distributed problem solving. As the Internet evolved rapidly, new computational systems emerged to emulate commercial systems. Solomon has developed demand-initiated self-organizing commercial systems for both intermediated and dis-intermediated transactions that employ novel multivariate and multilateral negotiation models.
One niche of the automated commerce system lies in the aggregation of autonomous agents. Precisely how to combine pools of autonomous agents for wholesale discounts presents an opportunity to remove a layer of distribution from commercial systems. This research stream is important because it provides clues as to how to develop coalitions of robotic agents for common purposes.
MRS models have been developed. The Nerd Herd is an example of an MRS using rule-based social behaviors for subsumption based foraging popularized by Brooks. Second, the Alliance architecture developed a modular approach to robot team behavior that includes inter-robot communication. Such communication allows for emergent cooperation. An additional version of Alliance (L-Alliance) accommodates the learning aspect of robotic agents in order to achieve a form of adaptation.
Arkin developed a “multiagent schema-based robotic architecture” in which team cooperation was modeled using a behavior-based approach without explicit inter-robot communication.
Dias and Stentz provide a market-based model for multirobotic coordination in which individual robots in a distributed environment negotiate with each other in order to agree upon a course of action. Such a model applies the contract-net protocol used with software agents in a distributed network to the robotics context for operation of groups of autonomous robots in dynamic environments.
Finally, Solomon developed a hybrid MRS model with military and industrial applications in which a hierarchical leader-follower approach is implemented in a hybrid central-control and behavior-based control architecture.
Most MRSs possess several common traits, including mobility, intelligence, communications, group behavior and specific functionality.
One critical aspect of robotic group behavior lies less in the value of intelligence that in the importance of methods of aggregation. It is a key challenge of robotic systems of determine ways for robotic agents to synchronize, cooperate and collaborate and, in sum, to work together as a team. The emergence of dynamic coalitions of robotic groups is one of the most interesting and important areas of robotic research.
The effort to achieve the development of complex MRSs that may emulate, and even transcend, emergent natural self-organizing processes, has become primarily a computation challenge that involves the need to create sophisticated AI architectures. AI systems have themselves emulated biological systems, with the advent, from Holland and Koza to the present, of genetic algorithms, genetic programming and evolutionary computation methods in order to solve complex problems. A related research stream involves A-NN, which has utilized GA in order to establish weight values of neural nodes. One main aim of the neural networks is to develop self-configuring and self-organizing learning systems for complex problem solving. This is useful in real time collective robotics situations in which rapid adaptation to a changing environment is necessary.
The development of hybrid AI technologies that synthesize various methods for specified problem solving would provide a robust and successful option in the computer scientist's arsenal of weapons that may be useful for the development of sophisticated MRS architectures.
The present inventions involve multi robotic systems, multi agent systems, collective robotics, artificial group behaviors, aggregation of robotic agents, coalition formation, dynamic coalitions, self-organization of robotic agents, emergent behavior of intelligent agents, cooperation of intelligent agents, multi agent learning, problem solving between conflicting intelligent agents, artificial intelligence, artificial neural networks and multi robotic operating systems.
Multi-robotic systems are complex networks that facilitate the interaction between autonomous robotic agents according to specific rules of behavior in order to perform a specific function or combination of functions. The present invention describes a system for multiple mobile robotic behavior by applying the logic of advanced computer science, in particular artificial intelligence (AI), with advanced robotic electronics and mechanics. The focus here is on artificial robotic collectives. So far very little research has been developed on the group behavior aspects of robotic societies as they plan, and then achieve, a coordinated goal.
There are several layers of any such collective robotic system, including (1) the computation, electrical and mechanical hardware of each autonomous robot unit, (2) a hardware network layer that links the individual robots together with wireless communications, (3) a metacomputing layer (that performs complex memory, database and computation analysis functions) in a node to node distributed computing model, (4) an omni-nodal artificial neural network (A-NN) layer for distributed AI, (5) an evolutionary A-NN layer—driven by genetic algorithms and genetic programming—for adaptive group learning in order to develop real-time cellular automata (CA) based simulations to seek optimal system solutions, (6) an OS layer and (7) a layer for specific functional applications.
The present invention describes a sophisticated MRS that is dynamic, interactive and evolving, adaptive to its environment and capable of exhibiting emergent behavior. The system is designed as a hybrid of behavior-based and central planning control processes in a distributed network environment. By decentralizing numerous functions in a distributed architecture model, groups of autonomous robotic agents can learn together, make group decisions together (cooperatively and competitively), negotiate and solve problems together, congregate together in various sub-sets and re-configure in non-overlapping sub-groups. Using these unique approaches, autonomous robotic agents can form and reform into various configurations of groups in a self-organized way interacting with each other and with the environment in order to achieve pre-programmed, or evolved, goal parameters.
Artificial intelligence (AI) is used in a number of MRS processes, including individual robot learning and decision making using genetic algorithms (GAs), genetic programming (GP) and other evolutionary computation (EC) approaches as well as group robotic agents that uses A-NN and hybrid evolutionary A-NN approaches (including GA, GP, FL, etc.) that provide tools for adaptive collective learning and decision making. The use of both individual agent and group learning tools are important because though the collective resources are far greater, when the system defaults to behavior-based biases, for instance, in situations with diminished computation resources, it is necessary for the individual robotic agents to have the tools to maintain autonomy. By building on the lower layers of behaviors of reactive approaches, a more complex MRS can evolve beyond ant society emulation.
In practical terms, MRS operation in unknown environments presents numerous challenges and problems to solve. In the absence of a centralized “mission control” station to solve all the problems a robotic system may encounter, there must be a number of fall back system modes in order for the mission to be successful, which leads to a hierarchy of system structures. These system modes are dependent on computation resources, communications resources, levels of robotic agent autonomy, levels of learning and levels of group behavior.
In earlier multirobotic systems, a relatively simple architecture would consist of a leader robot with various followers in a hierarchy. The leader possesses increased autonomy and orders the followers (super-drones). In this model, pre-selected squadrons are formed, the control for which can be manually intervened by human interaction processes such as a video feed for mission objective alteration. Reprogrammable orders and priorities can be uploaded at any time.
In one embodiment, supplementary external computation resources can be kept outside of the MRS and fed in as needed by satellite. Alternatively, though computation is performed externally to the MRS, analytical results can be used to control the system. In addition, reporting on agent behavior can be provided to an off-site blackboard so as to unify control at a central command center.
As the system and its agents gain autonomy, increased capacities are brought in, such as computation power, communication bandwidth and AI capabilities. Still, only reactive behavior-based autonomous robotic agent interaction approaches would yield a relatively simple system that appears to generate group behavior but merely mimics collective actions because of the outcome of interactions between autonomous agents. The system in this mode is merely semi-autonomous, which reflects its resource limits.
As the MRS system is linked together in a distributed network of autonomous robotic agents that employ powerful computation resources and AI processes, the system can automatically “think” like a group and constantly reconfigure to the best available situation while interacting with and adapting to its environment.
It is therefore valuable that the system, though using a hybrid architecture, employ a number of distinct embodiments that accommodate changes and that automatically default to the most complex task achievable.
Hybrid MRS Architecture with Distributed Resource Management and Command Structure
A pure behavior-based reactive MRS architecture has advantages of local control and emergent behavior but disadvantages of the inability to control large groups in complex adaptive environments. On the other hand, a central deliberative MRS control architecture has the ability to develop large self-organizing interactive systems and sub-systems but has the limits of being cumbersome and dependent on substantial computation resources. What is needed in order to build and operate a complex and high performance MRS is a hybrid architecture. In effect, the MRS architecture is a complex, continuously reconfiguring, operating system that links together robotic agents with computation, communications and software subsystems. Such a system must be modular (so that upgrades in a subsystem can be seamlessly performed), scalable (so that nodes can be added or removed) and reconfigurable. The system uses mobile software program code that provides inputs and outputs to robot machine agents. The “Harness” dynamic reconfigurable metacomputing model is a pioneer for this mobile self-organizing MRS hybrid approach because it continuously seeks to re-route the system to the optimal computation and communication pathways.
On a lower level, each robotic machine unit has sensors, actuators, microprocessors, communication receivers and transmitters, power supply, a specific functionality and (system and applications) software. However, when they are linked together, the opportunity exists for the MRS specific mobile robotic unit sensors to be organized into a network for collective data acquisition. The group's collective computation resources can analyze the sensor data. In addition, the group of mobile robotic agents can use complex AI induced learning processes to make group decisions, even in the face of noisy, error-prone and conflicting data streams. By maximizing the efficiency of the available group MRS resources, intelligent group behavior can emerge.
The aggregation of MRAs into subgroups can occur, further reconfiguring in complex ways in dynamic and changing environments. By learning and working as a group, specific autonomous agents are altruistic and may be sacrificed for the greater whole if it is necessary in a specific critical operation. Further, specific sub-groups may conflict and split the herd in order to achieve different objectives. The convergence of technologies that allows teams of autonomous MRAs to work together—computation resources and reconfiguration, communications bandwidth capacity and complex system software—make possible a revolution that emulates how groups of humans think and behave.
In order to make this technology convergence operability possible, it is necessary to develop a distinctive hybrid MRS architecture for a distributed self-organizing system. Such a hybrid system accommodates lower-level bottom-up reactive modular behavior-based approaches as well as the use of sophisticated hybrid AI resources (D-AI, A-NN, GA, GP, etc.) that work in a distributed system for group learning processes applied to complex decision processes, optimal simulation and collective robotics actions in dynamic environments. Such a hybrid model allows for adaptation in uncertain environments while also being able to carry out initial, and evolving, program objectives.
If one compares how animals work in groups we see a resemblance to our system. Though specific animals have sensory data, memory, navigation, data analysis, decision-making and action sub-system abilities, as a group collectives of animals can achieve marked performance improvements because they have more data and analytical capacities and the integration of successful actions that increase the probabilities of winning at foraging for food or defending against attacks. Why, then, cannot an MRS be developed that emulates, and even transcends, the performance of the animal (and insect) group model?
Historically, one of the main problems in building such an intelligent MRS of autonomous self-organizing MRAs has been computer resource constraints. There is the limit, not only of computer capacity, but also time, constraints. A huge amount of data must be processed in a short time while the MRS is operational; in essence, the system must compute on the fly as it gathers and understands data and decides what to do and then how to act as a group. There are practical solutions to these resource constraint problems. First, the application of Grid computing models provides an appropriate distributed model for maximizing computation capacity by sharing resources among MRAs in real-time. This model can be scalable so that new MRAs can be added as needed even if others are subtracted as the mission requires. In fact, each agent can be re-tooled and upgraded in each reuse of the modular system.
Second, multiple communications topologies can be used to re-route data streams to the most efficient use within the distributed system, including using advanced caching techniques for optimal collective effect. Finally, AI software can be employed for learning, negotiation, decision and simulation of complex collective behaviors. The system then determines, while it is mobile, what to do and then acts as a team to cooperatively achieve the objective. By overcoming the resource constraints with collective action, an intelligent MRS emerges.
The present system is therefore far more advanced than previous MAS approaches that seek to emulate the behavior of groups of simple insects because our system is endowed, not only with autonomous agent intelligence, but with collective group intelligence that transcends simple group behaviors. It is clear, then, that in order to develop such an advanced MRS, hybrid or meta-architectures must be employed that combine both local and global aspects.
Towards a Hybrid MRS AI Model: Distributed Problem Solving, Integrated Group Learning, Decision Processes and Dynamic Optimization Simulations with Cellular Automata
AI has emerged in the past generation as a valuable tool for solving complex problems. Genetic algorithms, developed by Holland and others, are a problem solving method to evolve, through reproduction, crossover and mutation techniques, algorithms. Genetic programming and other evolutionary computation approaches seek to solve different domains of problems. These complex strategies seek to emulate natural evolution processes so as to find the fittest, most efficient or optimal solutions.
The development of artificial neural networks (A-NN) was initially intended to emulate brain function. Referred to as connectionism, A-NN uses GA and FL (soft computing) techniques to map out, train and reconfigure a network of nodes for solving problems. By using an adaptive network architecture topology, the A-NN system can optimize adaptation to its environment. By training the network over distributed groups of agent nodes, the A-NN can learn. Evolutionary A-NN (E-A-NN), or neurevolution, is useful for reinforcement learning. A-NN's work by using genetic algorithms to adapt input features, learning rules and connection weights. One of the most effective applications for A-NN is nonlinear statistical models such as pattern recognition. A-NN's learn by altering synaptic weights; synaptic weight variables change by using fuzzy logic techniques to assess probabilities and thresholds. Bayesian networks use hypotheses as intermediaries between data and predictions to make probability-based estimates of solutions. Hopfield networks are used to remember an earlier network configuration and to revert to an old network when noisy data limits continuing network development.
The present invention uses a hybrid approach to AI that combines GA and GP with A-NN and D-AI architectures. The combination of evolutionary computation approaches with distributed neurocomputing models produces a system that constantly rewires itself as the system is reconfigured. This approach is necessary because finite computation resources need to be maximized even while the distributed mobile MRS changes. Not only is this scheme scalable but increased computation capacity can be provided on demand if needed by specific under used MRAs. Such a hybrid AI architecture is best suited for learning by groups in a distributed network as well as for optimal adaptation to dynamic environments.
Hybrid AI approaches can be useful when solving complex problems. Two main problem solving models involve either cooperative (altruistic) or conflict (self-interested) oriented agent behavior. One main computational challenge that involves MRS is the distributed problem solving that requires negotiation among conflicting autonomous agents.
Conflicting MRAs use AI approaches to negotiate a settlement so as to solve complex multilateral disagreements. One way for groups to solve problems in a conflicted MRS is by finding proper matches for shared common interests, thereby focusing on the limited remaining variables and disagreements. This pruning process can settle an issue either by pre-determined (or changing) rules or by a vote between involved agents. In this way teams of MRAs can compete for effective solutions. Another method to find solutions in conflicted MAS situations is to set up a competition for the strongest strategies according to agreed upon rules. Finally, an agent can persuade other agents to its position.
All of these models involve inter-agent collaboration for complex group problem solving. The resolution of competing rival MRAs conflicts result in agreement about an optimal solution. Through conflict and competition, not only is common ground sought, but a winning algorithm solution is determined for complex problems. This problem solving negotiation approach is useful for organizing heterogeneous MRAs for common objectives.
How are negotiations between MRAs in a MRS processed? Autonomous robotic agents use complex decision processes that ultimately affect group behavior. Decisions can be made by either individual agents or by groups of agents. Rules are used to prioritize specific possible choices over others. Upon achievement of a specific threshold, a decision process yields a resulting choice of possible options. Once a threshold is achieved, a plan of action can be implemented.
Since it is important to configure group decision processes for MRS problem-solving, a range of decision choice constraints present the lower and upper bounds of potential optimal solutions. Further, these parameters are constantly shifting in dynamic environments. Hence, methods need to be devised to find the shortest path to perform specific tasks. One way to do this is to perform specific tasks. One way to do this is to use statistical weighting to prioritize problems and solution possibilities. In the context of complex changing environments, an MRS must simultaneously work on solving numerous constantly changing problems. The Markov decision process makes decisions by prioritizing possible choice as measured by evolving values criteria.
MRS action starts with a plan. By mapping the parameters of group action plans, we can model the optimal configuration or allocate the most efficient resources. Decision logic processes lead to identifying trade offs (parameters) between possible solutions that lead to an optimal problem solving choice. MRAs use computation optimization techniques to select optimal solutions to complex problems in uncertain environments. By mapping various scenarios, using AI and decision processes in a distributed network, MRAs select the best plan to achieve objectives.
MRSs use advanced hybrid AI methods in order to achieve optimal grouping patterns of behavior. Unlike purely computational MASs, a MRS have physical dimension and motion in space. These physical and geometric realities about the practical operations of MRSs involve the need to organize spatial interactions and movements. It is useful to model these MRA movements before actually performing specific maneuvers primarily through the use of simulations.
Cellular automata (CA) models provide an important tool to simulate the changing movements of MRAs in an MRS. By using AI approaches, each robot is represented as a cell in a larger system. Cells can interact with neighbor cells in the neighborhood of a CA system, with two dimensional, three dimensional or four dimensional models representing the change in cellular states.
The results of combinatorial optimization approaches to seek the best solutions to solve problems can be represented by CA simulations and, thereby, tested, before actually implementing these decision choices. By modeling group behavior in real time, the MRS solves problems and can seek improved solutions that can capture subtle contingencies in complex operational situations. MAS swarms are tested in particle simulations using CA models, but MRSs have not applied these important CA driven simulations for real geometric behaviors. Therefore, the present invention uses simulations in a dynamic, rather than merely static, way, for real time testing. In the simulation, virtual robots are provided the valuable advantage of trial and error of potentialities of activity so as to learn from complex contingencies, in order to optimize the chances for mission success.
Swarms, Flocks, MRS Aggregation and the Formation and Reconfiguration of Dynamic Coalitions
Nature provides analogies for computer scientists in the contexts of AI and group robotics. In the case of AI, GAs and GPs seek to emulate natural selection by breeding the best fit problem-solving programs using principles of sexual reproduction, pruning and random mutation. In the case of robotic group behavior, scientists have sought to emulate insect (ant and bee) social behaviors in order to understand how compolex patterns emerge from simple individuals. How can MRSs be developed that have the self-organizing properties of insects? The two main behaviors that have intrigued observers are foraging (food location search and procurement) and swarming behaviors.
Scientists have discovered that ants use pheromones (chemicals that have an odor to attract others) to develop complex foraging behaviors. By laying down pheromones, which, though temporary, can be increasingly intense if compounded, ants provide a natural reinforcement mechanism (stigmation) with positive feedback. This positive reinforcement learning mechanism suggests a self-organizing system.
There are other ways for insects to communicate with each other. Some ants and bees have developed ways of communicating with their nearest neighbors about food sources, for instance, to get help with or altruistically share information with the group. This nearest neighbor communication approach, which is primarily sense based, is key to the formation of flocking, herding and schooling behaviors in animals and fish.
In the case of bees or ants, there may be specialists that perform specific functions in the hive or nest in order for the whole organization to function more smoothly. This division of labor has evolved for millions of years as an efficient social system.
Insects may communicate with each other indirectly. The process of stigmation operates with an insect affecting, or changing, the environment, which then catalyzes other insect behavior. The use of pheromones illustrates this process because the ants lay down an attracting chemical that may be acted upon by others in a limited time.
Animal and insect group behaviors emerge at the local level. Though insects are not intelligent in some ways, their complex group behaviors suggest that they have evolved social intelligence. By working in groups, they have defended against predators and survived in hostile environments. But here are limits to this kind of swarm intelligence.
Though they have an initiator, most swarm or flocking behaviors do not have a single persistent leader. Instead, such social behaviors focus on local and reactive interactions.
Flocking is a case in point. Each bird in a flock has limited information about flockmates. Instead, they have neighbors they provide local information on direction and speed. The big challenge is to avoid collision with neighbors even as they signal trajectory and velocity data through their behavior. Consequently, both attractive and repulsive forces are involving in flocking behaviors.
There are, then, simple flocking rules that are useful to MRS designers because they illustrate local reactive behaviors: (A) Fly at a steady state speed of neighbors, (B) Tend to the center of the flock and (C) Avoid collisions with neighbors. This is similar to driving on highways because we have limited information (visibility) restricted primarily to our nearest neighbors, with which we seek to avoid collision but also maintain a consistent pace. Flocking, like herding, school and swarm behaviors, have evolved to allow groups of insects, birds, fish or animals to move in a hostile environment while avoiding peripheral members from being picked off by predators. In nation, then, avoidance of obstacles, neighbors and predators has become integrated into the rules of survival that social group behaviors maintain.
How does a swarm form? An event will stimulate an individual insect to attract neighbors to the swarm activity. Though any individual can be a leader that initiates action, the recruitment of other individuals through attracting the cooperation of similarly interested neighbors is key to the process because these individuals then respond by attracting more neighbors, and so on. Thus, any individual can initiate a swarm or flock; this initiation is a sort of initial request to procure resources for a specific (defensive or offensive) function or activity. Rather than a centralized mission control issuing orders to the troops, specific decentralized individuals can trigger group activity in a sort of local reactive chain reaction process that has the effect of overwhelming an enemy. In some cases, specialists alone, such as soldier ants, may swarm for an attack process.
Insect and animal social behaviors are important to understanding complex social processes involving simple individuals. Attempts have been made to emulate biological system swarm intelligence for development of artificial systems of robots. For instance, Arkin's (1998) use of Brooks' simple modular reactive robot for group behavior shows an attempt to model complex behaviors from simple robots.
Beyond Flocking: MRS Aggregation and the Formation and Reconfiguration of Dynamic Coalitions
The present invention goes beyond these interesting biological emulations. Because our system is layered, with simpler default modes of operation, we will use simple swarm behaviors in an MRS that employ reactive local interactions. But the main objective is development of complex aggregated MRS systems that are capable of intelligent social behavior as well as the operation of dynamic coalitions. Whereas the simpler group behaviors have severely limited computation and communication resources in a homogeneous system, the present invention does not. Simple swarm behaviors have anonymous homogeneous simple members (in uniform roles) with primitive local communication, minimal computation capacity and the limits of reactive behaviors using a narrow set of rules for learning and action. The limits of this biologically inspired system can be improved by development of an advanced MRS that exhibits social intelligence. Our system has autonomous individual MRAs with highly advanced computation, AI and communications capabilities, complex learning and simulation functions, specialization features and team behaviors in a heterogeneous system. In short, the present invention emulates human social behavior by using artificially thinking mobile robotic agents for a range of functions.
The problem of how to aggregate objects is an important one in computer science. Methods of aggregation involve collecting together disjoint sets for an organized assembly. Combinatorial optimization is the mathematical field concerned with seeking solutions to aggregation problems. Aggregation is useful for mass pooling of customers with common interests for wholesale discounts. Similarly, combinatorial auctions are a useful commercial structure to enable parties to acquire bundles of items for optimum benefit.
For the purposes of the present invention, aggregation is important as a process for organizing groups of MRAs within an MRS. We are not only interested in how groups of intelligent robotic agents form, but also the process by which groups break into subgroups and reform. Intelligent aggregation of MRAs involves automatic selection, formation, combination, reformation and dissipation of groups. Each new set of intelligent agents represents a new configuration. Emergent behavior of the MRS leads to a complex self-organizing system that never settles on an equilibrium because it is constantly changing. Finally, unlike other pure computational contexts for aggregation, the application in an MRS involves the geometry of space and extension and the physics and mechanics of motion.
The autonomy of intelligent agents leads to the opportunity for individual specialization. Whether in biological or economic systems, specialization affords the optimization of teams because it establishes an efficiency enhancing division of labor. Groups of MRA specialists can work together in an artificial system for increased benefit to the objectives of the whole group. The existence of specialization also makes possible the interactions of sets of agents.
Aggregation is a process of grouping entities together. One useful way to model groups is with game theory. As applied to an MRS, game theoretic models have a geometric dimension. Game theoretic approaches to modeling an MRS is useful particularly because they can be multi-phasal and interactive. Not only are MRA interactions nicely modeled but complex interactions between sub-groups can be more optimally represented as well as interactions with the environment. Game theory can model cooperating agent behavior as well as conflicting or heterogeneous behaviors. An example of a heuristic for MRS game theoretic modeling parallels chess playing maneuvers, with openings, gambits and traps providing MRA models for the inter-operation of artificial societies. Robotic agents work together to develop winning game strategies for achieving goals or solving problems.
One of the aims of the present invention is to develop methods for MRAs to constantly develop shifting groups. We are interested in discovering how intelligent autonomous robotic agents form and reform into dynamic coalitions of collectives. Understanding precisely how sub-groups of MRAs organize, self-configure, reconstitute, adapt to their environment and regroup is the key to understanding complex emergent group behavior in intelligent self-organizing systems.
With severe resource constraints, mobile agents will tend to behave according to simple rules inspired from biological systems, with local and reactive control. But endowed with sufficient computation and communications resources, an intelligent MRS will be able to perform more effectively. One of the areas of improvement in the operation of groups of MRSs lies in establishing methods and processes for dynamic coalition behavior.
Multiple squads containing specialized MRAs can work together by sharing sensor data, data analysis, computation, communications and decision processes. Such multiple squads can form alliances and temporary coalitions for specific missions with numerous applications to industry. When group resources are restricted, specific squads can operate autonomously with limited information and still perform its objectives. Further, higher priority squads can get more resources at crucial times. Squads can reconstitute by taking resources from the larger group for continuous dynamic coalition reformation so as to more optimally adapt to changing environments. The existence of multiple micro-coalitions can be better suited to satisfying multiple goals simultaneously and thus increase the chances of a mission success.
Squads of MRAs break off from larger groups in an MRS. The squads can share the larger computation, communication and sensor resources and decision processes of the larger group. In effect, the squads operate as teams of nodes in a neural net that constantly reconfigures on the fly. Since some of the sensors in some of the squads are exogenous to each team, the squads have access to data streams beyond any limited team. Sub-teams are synchronized into the distributed network using hybrid AI approaches. Nevertheless, each squad, and its reconfiguring teammates, can work independently with local behaviors. In addition, differences between agents in a squad, for example, specialists or different “personalities,” can create complexity in squad behavior within the practical constraints of their programming, as they inter-relate in different configurations. Squads self-select into various coalition configurations, but during complex missions, new squads can pick up stragglers from previous damaged squads. Similarly, squads can merge in instances where combined strength is needed to solve a problem. Robotic agent nodes can be added or subtracted as the system continuously reconfigures to achieve optimum success.
Different methods are employed in order to realize group MRA self-organization processes. In one important sense, game-theoretic and cellular automata simulations are useful in order for collectives in an MRS to map out and achieve complex plans for problem solving. By employing these processes within AI driven computation, intelligent MRAs work together to optimize complex processes in order to achieve mission success. The opportunity to simulate these processes of constant re-grouping for dynamic coalitions of MRAs allows a new generation of applications of MRS social behaviors to be possible. In this way, among others, the present system far surpasses prior approaches to emulating biological social behaviors. Our system allows intelligent MRAs to constantly shift in dynamic coalitions that are best suited for environmental interaction. It is precisely the continuously changing environment that requires development of a complex system that makes possible continuous reorganization.
Innovations and Advantages of the Present System
The present system has a number of innovations and advantages over earlier inventions. These innovations involve (1) multi-robotic system architecture, (2) computation resource structure, dynamics and allocation, (3) AI dynamics, (4) group negotiation, learning and decision structures and processes, (5) intelligent social behavior involving mobile robots and (6) dynamic coalitions of MRAs.
The present invention utilizes a novel hybrid MRS architecture that dynamically adjusts from manual operation of groups of MRAs to wholly automated socially intelligent MRAs in order to accommodate severe resource restrictions as well as extremely complex behaviors. By defaulting to the most complex appropriate resource level, the system optimally adjusts to environmental conditions. For instance, very small MRAs may be resource constrained and would thereby employ simpler local reactive behavioral rules. The architecture of the present system is also both modular and scalable so that growth or shrinkage will not affect performance.
The present system uses a distributed wireless grid supercomputing model. This approach allows the sharing of computation resources, including memory, database storage and data analysis capacity, thereby far extending previous constraints. In addition, this distributed model is optimal for equal node parallel processing within a collective. Computation processing speeds of dozens of teraops could be maintained in this system, thereby providing ample resources for complex group behaviors. The present system also uses advanced routing procedures to maximize the most efficient geodesic heuristics.
The present system employs a novel use of a MAS within a MRS in order to communicate, negotiate, control and organize group behaviors. Intelligent mobile software agents (IMSAs) are the analytical representatives that perform critical internal functions in the robotic system. In addition, intelligent negotiation agents (INAs) represent a core and innovative aspect of the present system as a vehicle for MRAs to interact and solve problems.
The present invention uses a dynamic reconfigurable evolutionary A-NN that provides optimal adaptation to the changing environments of an intelligent MRS. The A-NN uses hybrid AI techniques, including combinations of GA, GP, FL and EC. As nodes are added or subtracted to the network, the A-NN is automatically rewired for maximum efficiency. The system uses feedback loops to learn. The A-NN is useful to train the system in group learning processes. These applications to a mobile and dynamic MRS are novel. The use of connectionism (neural nets) in a MAS and a MRS is a huge leap from earlier systems.
In order for the present system to learn, it employs FL processes that use probabilities to make group decisions by selecting the best available option among a range of contestant options. The system utilizes combinatorial optimization approaches to select the best solution to solve problems. Particularly in conflicting situations between agents, there is a need to negotiate a settlement by developing a method of winner determination. The system employs novel approaches to asymmetric problem-solving by using multi-lateral negotiation methods.
The present invention uses game theoretic approaches and cellular automata schemas in order to simulate tactical system opportunities for an MRS in novel ways. By using real-time CA and GT simulations, an MRS can automatically select an optimal problem-solving path and, hence, model complex interaction dynamics among MRAs and between MRAs and the environment. Given limited information in challenging environments with resource constraints, the use of simulation modeling for action planning and contingency scenario testing is necessary to achieve highly intelligent MRS behavior.
The present system is novel because it is heterogeneous. The MRS employs specialty robots for diverse functions. Some MRAs may have multiple functions, alternative functions or work in teams with complementary functions. This approach increases efficiency of task execution because it promotes an automated division of labor in an MRS.
Despite their specialty functions, any agent can initiate group behaviors. The attraction of MRAs to collectives can be demand-initiated in a novel implementation of group behavior in an MRS. This approach enhances system performance. In one implementation, stronger data inputs may constitute invitations to act beyond a specific threshold and thereby initiate MRA grouping behaviors. The present system uses novel group attraction initiation methods.
The present system synthesizes local control with deliberative planning. This hybrid architecture is novel and is possible only with the unique convergence of advanced computation technologies disclosed herein.
The present system uses novel approaches to dynamic coalition formation. Using these approaches, the MRS constantly reconfigures its structure and dynamics in order to adapt to environmental changes. This more effective adaptation provides increased speed, precision, efficiency and effectiveness in mission critical situations.
By applying distributed artificial intelligence approaches, the present system develops a way for groups of robotic agents to make decisions in cooperative and in conflicting situations in real time. This is a novel and important advance over earlier systems.
The present system implements novel MRS approaches involving tactical cooperating teams of MRAs. This sophisticated use of the system transcends earlier notions of artificial group intelligence.
Why are groups of robots important? Traditionally, robot groups allow an increased speed to do a task. Like in nature, groups are increasingly reliable since some may fail but the group still finishes the task. In addition, using robot groups to perform tasks can be more flexible than only individual robots. The present system offers higher performance benchmarks for these traditional advantages.
Since the present system uses multiple hybrid architectures, at the system and AI levels, there are nontrivial advantages over earlier systems.
The present system most efficiently implements complex group behavior in an artificial robotic system. For example, unlike earlier artificial systems that seek to emulate insect behaviors, the present invention seeks to emulate, and transcend, complex human group judgment to develop a true social intelligence. Consequently, the present invention goes beyond robotic systems that focus primarily on local control of the nearest neighbor and reactive behaviors.
The use of coordinated, cooperative and reconfiguring squads in dynamic coalitions in the present system provides numerous novel and useful advantages.
Finally, the present system is useful for a broad range of important applications, from manufacturing to toxic clean-up and from remote exploration to traffic coordination. The sheer breadth of collective robotic applications, to industry and beyond, using the present system suggests a range of uses that could provide revolutionary implications.
Applications of the Present Invention
There are numerous applications of the present system. Robots can have specific functions for specialized purposes. One robot can clean, while another can dry. But specialized robots can have particularly high utility as they function in teams. While specific purpose robots are useful, multiple function robots are increasingly productive. Multiple function robots can switch roles or change forms as needed to complete complex tasks. The more tasks a robot can do because of its multiple specialties, the more plasticity and flexibility it has.
Multi-functional teams of robots can perform more tasks than specific specialty robots. The more tasks that robots can do, the more plasticity of tasks a team of robots can perform because of the efficiency benefits of the maximized division of labor.
The following is an extensive (but not exhaustive) list of applications of groups of robots that the present invention advances.
The present system enhances factory production, assembly and distribution processes. Methods for groups of robots to work together may greatly accelerate production techniques. For instance, by using groups of multi-functional autonomous robots, a host products can be produced faster, more efficiently and cheaper than with earlier methods.
Regarding the factory assembly process, the novel use of groups of autonomous mobile robots may reshape the very idea of an assembly line because new interactive processes, reflecting an efficient modular workspace, will reconfigure approaches to activities in which parts are combined to a whole. The application of self-organized groups of multifunctional robotic systems to manufacturing assembly can promote just-in-time production processes and lean inventory to save time and increase efficiency.
The distribution function of factories, such as loading and unloading, can be improved with teams of autonomous robots working together. Such a system can replace routine labor practices.
Construction and Repair of Structures and Roads
Self-organizing teams of autonomous robots can build and repair roads and structures. From laying track or pipe to electrical, plumbing, framing and roofing, an MRS can be useful in performing laborious time-intensive routine structure building construction functions. Similarly, MRAs can be useful in the repair of buildings and streets. These novel MRA processes can save time and reduce costs of building construction as well as road work and repair. In one practical application, pot holes can be automatically detected and repaired by teams of MRAs.
Medical Applications: Medi-Bots
There are two categories of application of the present invention to the medical field. First, groups of medical robots (medi-bots) can be used in critical field situations to stabilize a patient. Autonomous medi-bots work together to (a) diagnose a patient's trauma, (b) resuscitate, via electronic pulse or CPR, a patient whose cardiac or pulmonary functions have ceased, (c) cauterize wounds to stop (or minimize) bleeding, (d) apply an IV for intravenous solution transmission in order to replace vital fluids and (e) call for more medical resources by providing a precise physical location position. Multiple medi-bots can much more efficiently rescue and stabilize patients, thereby saving lives.
Second, groups of medi-bot can assist doctors in clinical situations by performing functions typically attributed to nurses and assistants. Such medi-bots can monitor patient functions during procedures as well as actively support the surgeon or dentist so as to save time. These medi-bots can also supply expertise in critical operating room environments. In critical emergency room situations, where time and precision can make a difference, medi-bots can save lives.
Reconnaissance and Surveillance
A big category of use of the present system lies in reconnaissance and surveillance. Multiple autonomous robots working as a team are optimal for reconnaissance and surveillance activities. These MRAs can transmit real-time vision and sound to off-site locations, typically via satellites or terrestrial communications systems.
In one mode, the MRAs can be very small micro robots (more fully referenced below) that provide stealth advantages for reconnaissance and surveillance purposes.
In other embodiments, MRAs can be disguised as natural phenomena, such as animals, birds, insects, etc. for evasive and stealthy advantages. By emulating natural animal behaviors, mission effectiveness can be maximized.
Finally, by using groups of MRAs, a more complete, more dynamic and more accurate view of the terrain being viewed can be maintained than with any other existing technology.
If captured, an MRA in this system can erase its programming and be rendered a useless pile of sensors, while the remaining network nodes automatically reconfigure for effective performance.
Search and Rescue
Including reconnaissance MRAs and medi-bots, teams of robotic agents can conduct search and rescue operations in difficult terrain that may be inhospitable to humans, such as in extreme weather.
Groups of MRAs can be used to perform complex clean-up operations that may be hazardous to humans. These clean-up categories include: (a) toxic waste dumps, (b) nuclear reactor cleaning, (c) oil spill events and (d) sewer cleaning.
MRAs can use self-organizing maps of a local terrain to devise plans to most efficiently and safely provide toxic clean-up operations, thereby saving lives and protecting the environment.
Using similar configurations and methods as used in toxic clean-up applications of MRAs, an MRS can be used to fight fires. Ground MRAs can dig trenches and plot trajectories for the expanding fire territory, while aerial MRAs can drop fire retardant at tactical locations for optimal effect. As with toxic clean-ups, MRAs use self-organizing mapping processes to assess the scope and dynamics of the full-motion fire situation. Fire-fighting MRAs can save lives and protect property. This application can be useful for forest fires, urban fires or industrial structure fires that require complex problem solving and decisive action. Medi-bots can be used in conjunction with these fire fighting applications for maximum benefits.
MRA teams can be very useful for mining minerals in remote locations. Robots can identify the most promising locations to dig and then help with laborious digging and sifting tasks. Groups of MRAs can work faster and more efficiently than current automation processes, in part because they are mobile, autonomous and self-organizing.
Farming has enjoyed increased automation processes for generations so as to maximize production. Groups of MRAs can continue this automation evolution, particularly in the planting and harvesting contexts in which greater care is required for specific crops such as fruit and vegetables. In general, MRAs replace the routine functions of migrant pickers.
Like toxic clean-ups, ships have a number of complex and dangerous problems to solve. Because ships function as self-enclosed physical domains, MRAs can operate effectively on specific problems. Groups of MRAs can provide effective automated solutions to hazardous functions, thereby reducing risks and saving time and money.
One main activity for MRAs involves demining. Groups of autonomous robots can work together to either dis-assemble or explode mines that are discovered in a self-organized search process. In addition, disarming bombs can be a useful function for groups of MRAs.
Groups of automated vehicles can use the present system for effective operation. MRS vehicle categories may include cars, trucks, trains, aircraft and ships. In particular, cargo may be moved on various groups of autonomous vehicles for greater efficiency, timeliness and cost-benefit. Such traffic coordination systems may develop complex routing algorithms that emulate, and transcend, bird flocking or ant foraging behaviors.
Elevator and Dam System Regulation
Systems of elevators can be better guided and coordinated by using autonomous group logic. Similarly, dams can be regulated more efficiently by using group logic processes of an MRS.
The present system is useful to organize groups of weather balloons or aircraft to gather and disseminate data. The MRS is ideally suited to complex adaptive environments such as detecting dangerous weather conditions such as tornados or hurricanes. Groups of self-organizing MRAs can more rapidly predict dramatic weather system changes.
In an active mode, MRAs can not only predict poor weather but can influence its outcome. In a drought situation, MRAs can seed clouds to increase the likelihood of inducing rain. In an extreme case, MRAs can prevent tornadoes by influencing their movement very early in their development and changing the immediate environmental conditions. Only self-organizing groups of automated mobile robotic agents with specific functions—such as warming cool air in limited areas so as to retard or minimize a turbulent cyclic force—could execute this precisely or rapidly.
Groups of satellites can work together to perform distinctive functions such as optimally tracking moving objects by using the present system.
The present system can also be used to have groups of self-organized autonomous MRAs repair or readjust a satellite remotely.
As with other remote domains, the present system can be used in underwater applications. Specifically, the underwater context can be used with other applications, including surveillance, reconnaissance, search and rescuer and demining.
Remote and Space Exploration
The use of the present system for space and remote exploration is logical. By using teams of self-organizing MRAs, complex exploration activities can be routinely performed. This technology can be applied to underwater, extreme cold or deep space missions which are optimized for the flexibility and efficiency of the group behavior of mobile robotic vehicles. These vehicles can have multiple functions for the collection and analysis of local environmental data. In some situations, these MRAs can conduct covert operations during which they may need evasive programming capabilities.
Sentry Protective Services
Groups of MRAs can be used as an automated system of sentries for security protection purposes. Sentries can be used not only for surveillance but also for defensive uses in order to protect structures or personnel. Such MRA sentries detect and respond to invasive action by unauthorized personnel by tracking and evading the intruders and calling for assistance. In a more aggressive mode, automated sentries can respond to invasive behaviors by disarming and subduing unauthorized activities until the authorities can arrive.
The present system can be used by groups of MRAs that operate video or film cameras in order to capture dynamic movie scenes. Because the MRAs can be constantly moving and can be both self-organizing and synchronized, an MRA can facilitate a new generation of film-making techniques, particularly for the popular action sequences. While moving in synchronized or random ways, MRAs are well suited to capture moving scenes in distinctive cinemagraphic ways only possible in an MRS.
Commercial Laundry or Restaurant
Routine restaurant food preparation and delivery and commercial laundry functions can be done by teams of MRAs. Working as a group of short order cooks, MRAs can produce more variety of recipes in a shorter time than professional chefs or waitresses. Similarly, a commercial laundry service can be optimized by using groups of MRAs to organize, clean and package clothes. One hour discount cleaning is now possible by using an MRS.
One of the most exciting developments in robotics is the advent of small (fly-sized) robots. But the smaller the robot the greater utility is derived from working in groups. Once in groups, micro-robots can become very useful much as ants or bees are successful in groups. A number of group robotic applications involve the use of micro-robots. Given resource constraints of micro-robots, the group gains massive resource benefits while operating in a network using the present invention that make possible dramatic performance gains over merely a collection of unlinked autonomous robotic agents. These MRS micro-robotic networks could also be construed as very small scale integrated systems (VSSIS).
Generally, the smaller the micro-robotic agent, the simpler the system. Hence, some straightforward applications include surveillance and reconnaissance in which sensor data is transmitted for central use while the system is camouflaged as a natural phenomena (such as a real fly or spider).
Teams of self-organizing micro-robots utilizing local reactive operational behaviors can use traditional computer based group behavior that emulates biological system behaviors such as foraging or flocking. But the present system strives to go beyond these restrictive behaviors.
In one embodiment, disaggregated collectives of micro MRAs can form together into a larger composite robot exhibiting unified behavior. This is important so as to allow larger robots to disassemble into constituent (specialized) parts if necessary in order to evade a predator or disguise a maneuver.
In another embodiment, micro MRAs using the present system could inspect and assemble micro-electronic systems or could inspect biological entities for abnormalities.
As microprocessors progress to ever smaller sizes and greater capabilities, the practical uses and possibilities for micro-robotics, particularly in self-organizing groups, increases dramatically.
A close relative of micro-robotics is nanotechnology. The use by nanorobots and nanoprobes of the present system is logical. The same argument and restrictions of microrobotics apply to molecular sized nanorobotics. Like microrobots, nanorobots can assemble into larger composites that themselves work together as autonomous groups.
Uses of nanorobots include surveillance and reconnaissance. But more fanciful uses include biological applications that include cleaning arteries by injecting a group of nano-MRAs into a patient's blood stream. The nanorobots will go to the affected area, perform the operation internally and regroup for extraction. Nanorobots could also be used to identify and repair microelectronic abnormalities.
Groups of anthropological MRA “androids” can work together to form complex expert systems. Operating as consultants with autonomous opinions, robot expert groups can behave like specialist teammates to collect and analyze data, perform forecasting, develop alternative scenarios, make predictions and give advice in the form of reports. Such groups of expert consulting opinions can involve numerous substantive industry categories and topics, including optimal telecom and energy routing algorithms and economic, business industry and scientific analyses. The personalities, experience and learning processes of the android MRAs evolve. Taken together, such expert systems constitute a think tank. Ultimately, such a group of autonomous self-organizing robotic agents can form and reform coalitions of specialist experts similar to a sophisticated consulting firm. By applying evolutionary learning and combining various opinions, such complex systems can be creative and capable of original thinking approaches that far surpass chess playing supercomputers.
MRAs can take numerous forms. Since there are numerous applications of the present system in divergent industrial and technical contexts, it is appropriate to identify the structure and function of the variety of MRAs that can perform various jobs.
MRA vehicles can include various forms of aircraft, such as airplane, glider, helicopter, balloon, blimp, satellite or spacecraft. MRAs can operate in water as ships, boats, submarines or hovercraft. On land, MRAs can be automobiles, trucks, farm equipment, mining equipment, factory equipment, etc. There may be entirely new forms of MRAs as well, such as remote exploration devices, anthropological androids, micro-robots intended to emulate insect appearances, nano-robots and so on. The range of sizes and forms of MRA are very broad.
What unites the MRAs in the present system are common processes that make possible self-organizing group behavior of autonomous intelligent mobile robots. Nevertheless, the various specialized applications that are made possible by using the present system allow a broad range of important uses that endeavor to enhance the human condition by performing the riskiest, most remote, most complex, most routine and most important tasks imaginable.
Reference to the remaining portions of the specification, including the drawings and claims, will realize other features and advantages of the present invention. Further features and advantages of the present invention, as well as the structure and operation of various embodiments of the present invention, are described in detail below with respect to accompanying drawings, like reference numbers indicate identical or functionally similar elements.
List of Acronyms:
The present disclosures illustrate in detail the main ideas of the present system. Since the present invention has numerous embodiments, it is not intended to restrict the invention to a single embodiment.
The system and methods incorporated in the present invention are implemented by using software program code applied to networks of computers. Specifically, the present invention represents a multirobotic system (MRS) that includes at least two mobile robotic agents (MRAs). These MRAs have various useful purposes in the context of industrial and practical applications. The MRAs use complex software program code, including mobile software agents, to execute specific instructions involving robotic and computation operations. The software capabilities activate specific robotic functions within MRAs involving movement and decision-making.
The present invention focuses on how groups of autonomous MRAs operate in a distributed MRS. As such, the invention, or cluster of methods, solves problems in the area of computation for groups of mobile robots in a distributed network. The system shows novel ways for groups of MRAs to work together to achieve specific goals such as mapping the environment, coordinating missions, aggregating into dynamic coalitions and engaging in complex self-organizing activities. The system employs hybrid models for collective robotic control that combines not only synthetic control methods that combine central and behavior-based approaches but also hybrid artificial intelligence methods. Distributed artificial intelligence approaches are used in several contexts of the present system, including learning, negotiation, simulation and decision-making of MRAs and intelligent mobile software agents (IMSAs).
The main approach for decision making of MRA collectives is decentralized. In order to achieve self-organizing aggregation for specific missions in changing environments, the MRS engages in learning and decision processes that employ extensive use of IMSAs. IMSAs interact with each other to handle routine matters between MRAs, including communication, analysis and negotiation. Intelligent negotiation agents (INAs) provide a medium for multilateral interaction of MRAs for group decisions. Simulations are used extensively to model and select optimal pathways for MRA group action and for the evaluation of scenarios for action.
This detailed description of the figures is divided into several parts that explain: (1) the main structure and operation of the MRS, (2) resource management of a distributed MRS, (3) MRA learning, (4) AI and ANN, (5) IMSAs, (6) INAs, (7) problem solving, (8) cellular automata (CA) simulations, (9) aggregation and self-organizing dynamic coalitions and (10) specific applications including (a) remote sensing, (b) hazard management and (c) building processes.
General System Architecture and Dynamics
The second layer is the level of the mobile robotic system in a distributed network which connects together individual MRAs using communications. The Grid computing architecture is used to link the MRAs together at layer three in order to share computation and database resources between the individual MRAs for maximum network efficiency. In this way, the MRA network develops dynamic clusters for optimal computation and storage capability. Particularly in time constrained dynamic environments, the mobile Grid network model is critical in order to accomplish complex tasks.
At level four, the dynamic distributed database system is used. This extension of the Grid computing hardware architecture uses object relational databases and temporal data objects to organize data between databases in the MRAs.
Artificial intelligence is used in layer five as a dynamic interactive artificial neural network that evolves. By applying AI to evolving networks of MRAs as they interact in a dynamic environment, complex learning and adaptation processes develop.
Intelligent mobile software agents (IMSAs) operate within the multi-agent system (MAS), comprising the sixth layer. The IMSAs are complex agents that perform a number of important functions within each MRA, such as analysis and decision-making, and between MRAs, such as data search, negotiation and collaboration.
The MRAs produce complex simulations to represent their relative positions and movements as well as to map out the possible scenarios for future action. These simulations are represented as mobile cellular automata in level seven.
Finally, the specific functional application of each implementation of the system comprises level eight. The main application categories of remote sensing, hazard management and manufacturing processes each use specific functional representations that are closest to the environment with specific hardware types.
Layer three illustrates several main hybrid control systems that combine both central planning and behavior-based control models: (1) planning driven (0220), (2) advice mediation (0230), (3) adaptation (0240) and (4) postponement (0250). The planning-driven approach to combining the main control methods determines the behavioral component; it is primarily a top-down model. The advice mediation approach models the central planning function as advice giving, but allows the reactive model to decide; it is primarily a down-up model. The adaptation model uses the central planning control module to continuously alter reaction in changing conditions. Finally, the postponement model uses a least commitment approach to wait to the last moment to collect information from the reactive control module until it decides to act.
Finally, at layer four, the suite of synthetic control systems (0210) is constructed of various combinations of these main hybrid control models. For instance, a robotic unit may use a suite of hybrid control systems in order to optimize specific situations.
The evolution of these hybrid control models, as represented in the layered structure of figure two, is suited to complex social behaviors of a distributed MRS used in dynamic environments.
The structure of the dynamic database organization is referenced in
Taken as a whole system of MRAs in a distributed network, the MRAs manage data within a network and share database organizational functions. Similarly, in the distributed network, the software agents become mobile and interact with other software agents at various MRA locations. Finally, on this network level linking MRAs, computation resources are constantly restructured so as to maximize computer power for complex time constrained applications.
On the level of mobility, MRAs change spatial positions in variable time sequences in order to perform specific tasks. The software agents are also mobile within a limited wireless range between mobile MRAs. The network of MRAs constantly rewires its computation resources by using AI and ANN in order to adapt to its environment and to optimally perform the collective mission.
An individual MRA can detect an object (0520) in the environment with its sensors and change position from 0520 to 0530 as illustrated in
The individual autonomous MRAs are part of a distributed network in much the same way that inert computers are linked together into grid computing networks for supercomputing. This mobile grid computer network comprised of individual MRAs uses wireless communications in order to share computation resources.
MRAs possess not only computation capability, which allow up to teraops (one trillion operations per second) or yodaops of system processing power, but also database storage capacity as well. Each MRA possesses a database. However, taken as a whole, the MRS network comprises a distributed database system with complex coordination capabilities. The databases work together to store data objects such as a table, a calculation, a multimedia segment or other complex combinations of coherent mobile code. Such working together involves sharing database storage among a number of machines in order to ensure maximum efficiency under severe time constraints.
One of the particular types of objects that the object-relational database management system organizes involves temporal objects. Because the MRS is typically time constrained in order to perform its primary missions, temporal objects become a prominent part of the distributed database system. Temporal objects reveal their temporal priority in order to be listed in a higher or lower relative priority in the database for storage retrieval purposes. Objects are “tagged” with temporal priorities such as “now”, “imminent”, “very soon”, “in the future”, “possibly useful in the future”, “past”, “near past”, “immediate past”, “urgent priority”, etc. By storing, and reprioritizing, objects according to temporal priority, the system can operate much more efficiently.
The query generator (1010) requests the query executor (1020) to access databases at DB1 (1030) and DB2 (1050) in sequential order. These databases access the data object (1060), which is tagged as it undergoes temporal change and is given temporal priority (1040) and is then provided back to the query executor (1020). Once again the databases are accessed with temporal information about the data object. The data object is then directed to the query manager (1070) for feedback to the system. By prioritizing data according to temporal priority, the system can route data efficiently and effectively anticipate functions. Temporal data is useful in the present system in the context of evolving learning, evolving ANN, evolving game theoretic negotiation applications, evolving environmental conditions and general systemic adaptation processes.
It is necessary for MRAs to obtain and transmit information from other MRAs about specific data such as physical position, analysis, negotiation and decision-making. Concise data sets are transmitted between MRAs in real time about the location and analytical state of the MRAs. These abbreviated data sets are consolidated in each MRA by autonomous blackboards, which act as “radar readouts” informing MRAs about the state of the network.
MRA attitude biases are shown in
Environmental interaction is critical for learning and adaptation.
The MRA training process includes a combination of environmental interaction with group sensor data as illustrated in
The combination of the aforementioned learning approaches present a hybrid learning model with time constraints illustrated in the flow chart of
Given the distributed environment of the present MRS network and the learning schema presented, it is possible to have asymmetric MRA leaders. That is, if this is not a centralized system, it is still possible to have mission leaders, but they are not necessarily centralized or even consistent. Like in a flock of geese, any member of the flock may be a leader, though temporarily. Consequently, asymmetric MRA leadership provides the emergence of temporary hubs of MRAs that cluster together to interact with the environment.
A division of labor can occur in specialized teams for increasingly efficient performance as shown in
One key application of the (social) learning and environmental interaction processes is to construct a self-organizing map of an uncertain environment. This map can then be used as a benchmark for further collective action.
Using artificial intelligence and artificial neural networks optimizes the learning process.
In his quest for software that would solve complex optimization problems, Holland sought a solution from nature. By emulating biological processes of breeding, mutation and survival of the fittest, he sought to develop a new kind of software logic that would automatically improve in order to solve problems. His revolution in software design emerged as genetic algorithms that are binary representations of genes that undergo evolutionary processes similar to biological entities.
Holland's student, Koza, developed genetic programming based not on binary algorithms but on the evolution of trees diagrams.
The process of producing multiple generations of algorithms may take an enormous amount of time because there may be many thousands of generations before a solution to a problem emerges. In order to abbreviate this process, the evaluation part of the process may be performed in a parallel way. By breaking down the fitness evaluation function, the process is expedited.
Genetic algorithms, genetic programming and evolutionary computation techniques are applied to artificial neural networks in order to (1) calculate the initial weight and the connection weights of the signal between neurons, (2) train and optimize the connection weights, (3) generate the architecture and topology of a NN and (4) analyze the pattern, structure and phase state of a NN. GA, GP and EC are also applicable to a range of complex computation problems, including (1) distributed problem solving, (2) group learning, (3) group cellular automata simulations, (4) routing of computation resources in the distributed system, (5) scheduling in a dynamic distributed system, (6) creating a self-organizing map, (7) solving optimization problems, (8) performing game theoretic simulations, (9) performing parallel data mining and (10) selecting a winner from among complex aggregation choices.
ANNs are parallel computational systems including interconnected nodes. Sometimes called connectionism because of an emphasis on the connections between the nodes, ANNs have inputs and outputs in the connection weights between nodes. An ANN node represents an artificial neuron that is modeled after biological neurons in a brain. A perceptron is the structure that represents the sum of a neuron's inputs and outputs.
In phase one, the multi-layer ANN has inputs that register higher relative numbers at the top (connections between 3315 and 3310, between 3310 and 3320 and between 3325 and 3320) of the network than at the bottom (connections between 3315 and 3325, between 3325 and 3330 and between 3325 and 3320) of the network. As the network grows, shown in phase two, it emphasizes growth at the top, where there is significantly higher activity, and adds nodes at 3340 and 3350, while lower positioned nodes at 3355, 3360 and 3365 become inactive.
Fuzzy logic is a method to provide new approaches to computing that includes terms like “maybe,” “possibly” and other partial and soft descriptions. Also called soft computing, FL represents a departure from traditional hard computing with mutually exclusive logic. FL uses statistical methods to compute solutions to complex real world problems. FL is applied to ANN to produce complex adaptive networks.
A multilayer neural fuzzy inference network is illustrated in
One of the advantages of applying evolutionary computation to ANN is that such advanced computing can be performed more efficiently by using parallel approaches to break down a problem into smaller parts so that a larger number of computer processors may solve the problem simultaneously. In this way, multiple MRAs may work on a problem together in order to accomplish the task in real time. One application of this approach is in the fitness evaluation part of the genetic algorithm population production process. The problem of identifying the successful candidates in a population can be performed, and expedited, by using parallel processing.
In another example of using parallel processing to accelerate computation problem solving, consider the problem of adding a neural node. By using parallel computation processes, not only can the neural node be added, but the connections to it can be added simultaneously; in addition, the architecture of the network can be configured and reconfigured in real time, as new training models are considered and tested. The application of parallel algorithms to evolutionary computation, and, in turn, the application of EC (both GA and GP) to ANN provide increasingly efficient approaches for use in a distributed mobile MRS. MRAs share ANN computation in a flexible way as illustrated in
The application of EC and ANN to a distributed mobile MRS involves several important areas, including learning, training, adaptation and prediction. In order for MRAs to interact with an uncertain, and changing, environment, it must learn, predict and adapt. While EC is useful to train ANNs, it is the general learning capabilities that are regarded as an outcome of this training process that ANNs ultimately provide to MRAs and to the MRS that is critical to the effective real time adaptation needed by the system. Many of the problems that a mobile MRS encounters involve evolving solutions, adaptive behavior patterns, complex predictive scenario modeling and self-organized processes. These problems are solvable by applying EC and ANN models.
An example of the application of EC and ANN to an MRS is the modeling of game theoretic interactions. A particular strategy may be evolved for a particular player based on a basic rule pattern selection organized by a multilayered feed forward perceptron. Each layer performs a calculation of the weights of inputs, connections and biases. A random number of nodes is selected in the multilayer network, with a random number of offspring replicated from each parent and randomly mutated. A number of rules of game moves are identified and consistently applied. Each network generation is evaluated for accurate effectiveness of achieving a successful game move. The network is trained and retrained with full information. In this way, the learning process is refined so that each player is able to optimally move according to the rules.
This straightforward application of EC and ANN to a game theoretic modeling problem is relevant for a distributed mobile MRS because the present system uses simulations to model action. The simulations, which are discussed below in
IMSA dynamics within the MAS are discussed in
Search agents are described in
The general negotiation process is described with reference to intelligent negotiation agents (INAs) in a distributed network in
Because INAs are used in a critical way in a distributed mobile multi-robotic system, they are further developed in
INAs work by negotiating between at least two MRAs. INAs use argumentation methods to negotiate by presenting arguments with variable weights. INAs also negotiate about the best simulation to use in a specific situation. In general, INAs use multi-lateral and multivariate negotiation in order to come to agreement between noncooperating MRAs. In the case of competitive MRAs that negotiate for a compromise, problems are solved using group problem-solving and analytical techniques. Solutions to complex MRA group problems include the optimal or a temporary choice between solution options. Group problem solving is discussed in FIGS. 59 to 67. In all cases, AI is used in order to facilitate the negotiation and problem solving processes.
In reference to
The agents engage in (5330) and complete (5333) negotiation tasks, after which the initiator INA notifies its home MRA of its remote location activities by sending a message (5336). After reviewing more tasks at the remote INA 2 location (5340), the initiator INA either terminates (or returns home) (5343) or assesses additional tasks using internal database and analysis (5347), assessment (5350) and identification (5353) of the next location for task execution and moves to another location (5356).
After moving its program code (5360), the initiator INA identifies a need for AI computation (5363), requests AI computation resources at a specified location (5367), identifies available AI computation resources (5370) and messages a request for AI computation resources to be sent to a specific location (5373). The initiator INA receives (5377) and tests the AI computation resources at a specific negotiation site (5380). The negotiations are completed at the remote location (5385) and the initiator INA returns home (5390).
As shown in this figure, though a one-to-one interactive negotiation is shown between an initiator INA and another INA, an initiator INA (or its micro-agents) may negotiate simultaneously with at least two INAs at two or more INA locations in another embodiment.
The argumentation process is shown in
Negotiation is a process that fits into the overall game theoretic model that organizes competitive agents across limited goods. In this sense, negotiation involves agent strategies that anticipate opposing agent strategies.
Solution options between MRAs are described in
There are times when an optimal solution to a problem is not possible. In these instances, the best we can hope to achieve is the best available solution in a specific circumstance.
Clearly, the time aspect of the decision process is important because perfect information is rarely available and because agents in a multi-robotic system that interact in uncertain and dynamic environments benefit from waiting for the latest available information before deciding to act.
The group problem solving process requires specific analytical methods, including multivariate analysis, regression analysis, trend analysis and pattern analysis, in order to select a successful candidate.
Much of the substance of the problem solving, and negotiation, processes underlying inter-MRA conflict involves simulations. Because MRAs are mechanical entities that assume physical shape and mobility in space and time, it is possible to model them by using simulations. The MRS may use a number of types of simulations, including cellular automata simulations, particle simulations and game theoretic simulations. All three main types of simulation add valuable qualities to the representation of complex activities in a mobile distributed multi-robotic system, including structuring the dynamics of aggregation processes.
Cellular automata (CA) is a system of cells which are represented digitally as a binary unit or vacuum. As objects move through a grid, they fill up the space in the cell. If an object does not occupy a cell, it is empty. In this straightforward way, CA can simulate groups of objects in space and time. CA's may include two dimensional, three-dimensional or four-dimensional (i.e., including the time dimension) structures. Once including the time dimension, it is possible to model CA simulations. CA simulations are well suited to represent mobile distributed multi-robotic systems because the MRAs are seen as merely objects that move in space and time across a map in a nonoverlapping environment. Though the simulations may be complex, for instance, in modeling dynamic coalitions in adaptive sequences as they interact with a fast changing environment, their representation is critical in order to provide a mechanism for the self-organization of the MRS processes.
The initiation of homogenous MRA group formation is described in
The first phase of a demand-initiated environmental adaptation is described in
In FIGS. 79 to 84, dynamic coalitions, or re-aggregation processes, are discussed. In
The parallel dynamic traveling salesman problem is described with cooperating autonomous agents in
The general dynamic coalition process is described in
The current system is also applicable to cinematography, wherein one mobile object (or cluster of mobile objects) are sensed and tracked with MRAs. This process is described in
In a similar way as cleaning up toxic spills on land,
Roads can be built by using multiple MRAs as illustrated in
It is understood that the examples and embodiments described herein are for illustrative purposes only and that various modifications or changes in light thereof will be suggested to persons skilled in the art and are to be included within the spirit and purview of this application and scope of the appended claims. All publications, patents, and patent applications cited herein are hereby incorporated by reference for all purposes in their entirety.